What GenAI actually changed about Shopify development

Two years of building Shopify work with AI in the loop, and the honest summary is this: it collapsed the part of the job that was typing, and barely touched the part that was thinking.

What genuinely got faster

Liquid and section schemas. Theme sections are repetitive — a schema block, some settings, a loop, responsive classes. This is the single biggest time sink in custom theme work and it is now largely a generation problem.

API glue. Admin GraphQL queries, webhook handlers, Storefront API fragments. Structured, well-documented, low-judgement work.

Migrations and refactors. "Convert these twelve sections to the new settings schema" used to be an afternoon. It is now a review task.

Reading other people's code. Inheriting a theme somebody else built is a normal part of agency work, and comprehension speed went up sharply.

What did not move at all

Knowing what to build. A merchant says "the cart isn't converting". Working out that the real problem is the shipping threshold being invisible until the last step is not a generation task.

Data modelling. Whether a product family should be variants, linked products or metaobjects determines everything downstream. Get it wrong and no amount of fast code saves you.

Performance. Something that generates working code will happily generate working code that ships 400KB of JavaScript to a product page.

Anything involving money. Tax logic, checkout behaviour, subscription billing. The failure mode is silent and the blast radius is the merchant's revenue.

The commercial consequence

If most of your value was implementation speed, that value is depreciating. The work that holds its price is the work where being wrong is expensive: data architecture, integrations that touch payments and logistics, and knowing which of five plausible approaches survives contact with the merchant's actual catalog.

I now quote differently. Theme build timelines came down and I passed that on, because pretending otherwise is not sustainable when the merchant can see what tooling exists. What went up is the share of a project spent on discovery and on the integration surface — the parts that were always the risk, and now are visibly most of it.

A working rule

Let it write anything you could check in under a minute. Write anything yourself where a subtle mistake would take a week to surface. That line moves over time, but it moves slower than the hype suggests.